[BBVectorize] Don't vectorize selects with a scalar condition and vector operands.

This fixes PR27879.

Differential Revision: http://reviews.llvm.org/D20659

llvm-svn: 270888

2 files changed, 41 insertions, 1 deletions

diff --git a/llvm/lib/Transforms/Vectorize/BBVectorize.cpp b/llvm/lib/Transforms/Vectorize/BBVectorize.cpp
index 72ac7a5ba2b..af594cb751a 100644
--- a/llvm/lib/Transforms/Vectorize/BBVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/BBVectorize.cpp
@@ -886,9 +886,16 @@ namespace {
       Type *DestTy = C->getDestTy();
       if (!DestTy->isSingleValueType())
         return false;
-    } else if (isa<SelectInst>(I)) {
+    } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
       if (!Config.VectorizeSelect)
         return false;
+      // We can vectorize a select if either all operands are scalars,
+      // or all operands are vectors. Trying to "widen" a select between
+      // vectors that has a scalar condition results in a malformed select.
+      // FIXME: We could probably be smarter about this by rewriting the select
+      // with different types instead.
+      return (SI->getCondition()->getType()->isVectorTy() == 
+              SI->getTrueValue()->getType()->isVectorTy());
     } else if (isa<CmpInst>(I)) {
       if (!Config.VectorizeCmp)
         return false;
diff --git a/llvm/test/Transforms/BBVectorize/vector-sel.ll b/llvm/test/Transforms/BBVectorize/vector-sel.ll
new file mode 100644
index 00000000000..cb775ceae69
--- /dev/null
+++ b/llvm/test/Transforms/BBVectorize/vector-sel.ll
@@ -0,0 +1,33 @@
+; RUN: opt < %s -bb-vectorize -S | FileCheck %s
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+@d = external global [1 x [10 x [1 x i16]]], align 16
+
+;CHECK-LABEL: @test
+;CHECK: %0 = select i1 %bool, <4 x i16> <i16 -2, i16 -2, i16 -2, i16 -2>, <4 x i16> <i16 -3, i16 -3, i16 -3, i16 -3>
+;CHECK: %1 = select i1 %bool, <4 x i16> <i16 -2, i16 -2, i16 -2, i16 -2>, <4 x i16> <i16 -3, i16 -3, i16 -3, i16 -3>
+;CHECK: %2 = shufflevector <4 x i16> %0, <4 x i16> %1, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+;CHECK: %3 = shufflevector <4 x i1> %boolvec, <4 x i1> %boolvec, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+;CHECK: %4 = select <8 x i1> %3, <8 x i16> <i16 -3, i16 -3, i16 -3, i16 -3, i16 -3, i16 -3, i16 -3, i16 -3>, <8 x i16> %2
+define void @test() {
+entry:
+  %bool = icmp ne i32 undef, 0
+  %boolvec = icmp ne <4 x i32> undef, zeroinitializer
+  br label %body
+
+body:
+  %0 = select i1 %bool, <4 x i16> <i16 -2, i16 -2, i16 -2, i16 -2>, <4 x i16> <i16 -3, i16 -3, i16 -3, i16 -3>
+  %1 = select i1 %bool, <4 x i16> <i16 -2, i16 -2, i16 -2, i16 -2>, <4 x i16> <i16 -3, i16 -3, i16 -3, i16 -3>
+  %2 = select <4 x i1> %boolvec, <4 x i16> <i16 -3, i16 -3, i16 -3, i16 -3>, <4 x i16> %0
+  %3 = select <4 x i1> %boolvec, <4 x i16> <i16 -3, i16 -3, i16 -3, i16 -3>, <4 x i16> %1
+  %4 = add nsw <4 x i16> %2, zeroinitializer
+  %5 = add nsw <4 x i16> %3, zeroinitializer
+  %6 = getelementptr inbounds [1 x [10 x [1 x i16]]], [1 x [10 x [1 x i16]]]* @d, i64 0, i64 0, i64 undef, i64 0
+  %7 = bitcast i16* %6 to <4 x i16>*
+  store <4 x i16> %4, <4 x i16>* %7, align 2
+  %8 = getelementptr [1 x [10 x [1 x i16]]], [1 x [10 x [1 x i16]]]* @d, i64 0, i64 0, i64 undef, i64 4
+  %9 = bitcast i16* %8 to <4 x i16>*
+  store <4 x i16> %5, <4 x i16>* %9, align 2
+  ret void
+}